TL;DR

OpenAI has released ChatGPT for Teens, a restricted version that under-18s are placed into automatically — either because they say they are 13 to 17, or because the company’s age-prediction system estimates they are. It bundles study tools with tighter content limits and parental controls, and applies an under-18 model specification that rules out romantic language and emotional dependence.

What teenagers actually get

The learning side is the pitch. Study Mode answers with guiding questions and step-by-step scaffolding rather than a finished solution, and a new homework reminder tries to spot a teenager shortcutting an assignment and push them back into it. Quizzes and learning visualisations sit alongside. Study Hours lets a teenager or a parent schedule windows where Study Mode is on by default.

OpenAI also announced a partnership with CodeAI aimed at teaching how the technology works rather than just handing it over. Its chief executive, Karim Meghji, framed the goal as students being able to “catch its mistakes, and know when to stop trusting it” — an unusually deflationary line from a company selling the thing.

The restrictions

Content protections are on by default rather than opt-in, covering eating disorders and self-harm, sexual or graphic material, violence, and activities that put a teenager in danger. Parents with a linked account can set Quiet Hours, adjust some settings and receive alerts in narrow high-risk cases, now extended to eating disorders.

The model spec goes further than blocking romantic roleplay: the assistant is not supposed to use romantic language, foster emotional dependence, or suggest it has feelings or consciousness. Break reminders and persistent cues that this is software, not a person, round it out. OpenAI has added under-18 evaluations to its system cards so the claims are at least measurable.

Looking forward

For UK schools and parents the relevant frame is the Online Safety Act, which puts duties on services likely to be accessed by children. A vendor-built teen tier is not the same as regulated age assurance, and the whole structure rests on age estimation — a system that infers, and can be wrong in both directions. Britain’s age-assurance apparatus has had its own credibility problem recently: a UK certifier admitted using ChatGPT after errors turned up in its citations. Schools deciding what to sanction will want to know how often the estimate misses before treating this tier as a safeguard.